-rum - rensselaer polytechnic institutehomepages.rpi.edu/~zhaoz6/docs/mixrum_poster.pdf · k-rum...
TRANSCRIPT
LEARNING MIXTURES OF RANDOM UTILITY MODELSZhibing Zhao, Tristan Villamil and Lirong Xia
Computer Science Department, Rensselaer Polytechnic Institute
RANK AGGREGATION
• Alternatives: {Donald Trump, Hillary Clinton,Gary Johnson}.
• Mechanism: plurality rule.• Decision: Donald Trump.
RANKING MODELS
• Set of alternatives: A = {a1, a2, . . . , am}.• Parameter space: Θ.• Sample space: i.i.d. rankings over A.• Distributions: {Prθ : θ ∈ Θ}.
RANDOM UTILITY MODELS (RUMS)• Sample a random utility for each alternative independently.• Rank the alternatives w.r.t. these utilities.• The Plackett-Luce model: all utility distributions are Gumbel distributions.
MIXTURES OF RANDOM UTILITY MODELS (k-RUMS)• Parameter: mixing coefficients (α1, α2, . . . , αk); RUM components: RUM1, RUM2, . . ., RUMk.• Sample a component r w.r.t. mixing coefficients.• Sample a ranking from component r.
MOTIVATION: MODEL FITNESS ON PREFLIB DATA
Given the number of parameters d, the number ofrankings n and the likelihood of the estimate L:• AIC = 2d− 2 ln(L).• AICc = AIC + 2d(d+1)
n−d−1 .• BIC = d ln(n)− 2 ln(L).
Observation:
k-RUM � k-PL � RUM � PL,
where A � B means that the number of datasetswhere A beats B is more than that where B beats A.
THEORETICAL RESULTS: IDENTIFIABILITY
Def: For all ~θ1, ~θ2 ∈ Θ, Pr~θ1 = Pr~θ1 ⇒~θ1 = ~θ2.
Theorem 1. Let M be any symmetric RUM from thelocation family. When m ≤ 2k − 1, k-RUMM over malternatives is non-identifiable.
Figure 1: The label switching problem
Theorem 2. For any RUMM where all utility distributions have support (−∞,∞), when m ≥ max{4k − 2, 6},k-RUMM over m alternatives is generically identifiable.
FIRST ALGORITHMS TO LEARN k-RUMGeneralizd Method of Moments (GMM).
1. Choose q events: all pairwise comparisons andselected triple-wise comparisons.
2. Compute the empirical probabilities b1, . . . , bq .3. Find the parameter that minimize
∑qi=1(bi −
pi(~θ))2, where pi(~θ) is the probability computed
from the model.
E-GMM.
1. E-step: compute the probabilities that a rankingbelongs to each component.
2. M-step: compute the parameter of each compo-nent using the GMM algorithm by Azari Soufiani(2014).
The Sandwich Algorithm (GMM-E-GMM).
1. Run GMM for a good starting point.2. Run E-GMM to improve accuracy.
Figure 2: The sandwich algorithm
EXPERIMENTS
Figure 3: 2-RUM over 6 alternatives Figure 4: 4-RUM over 15 alternatives
COMPARISON WITH k-PLS
k-PLs k-RUMsClosed-form likelihood? Yes NoProving identifiability? Hard Harder
REFERENCES
Hossein Azari Soufiani, David C. Parkes, and LirongXia, “Computing Parametric Ranking Models viaRank-Breaking". ICML-14.Zhibing Zhao, Peter Piech, and Lirong Xia, “LearningMixtures of Plackett-Luce Models". ICML-16